Skip to main content
The policy in one sentence: Prem API does not store your prompts, completions, or audio after it returns the response. The platform keeps zero inference content.
This page tells you about confidential inference. The Zero Data Retention mode gives the same result, but the basis is different. Confidential inference stops retention with cryptography and hardware. ZDR stops it with a contract and with operational controls. See ZDR security boundary.

​
What Prem API never stores

The platform does not retain inference content:
  • Prompts and messages: not stored after the response
  • Completions and model output: not stored after the response
  • Audio for transcription and translation: not stored after the response
The encrypted inference endpoints are named RVENC: raw volatile encrypted. Your content exists inside the enclave only in volatile memory, for the duration of the request. See How It Works for the request lifecycle.

​
What Prem API stores

Some data is necessary to operate the platform:
  • Files that you upload: The platform stores files that you upload to the file endpoints. The files are encrypted with your keys. Only you can decrypt them. You can delete them at any time. See Encryption.
  • Request metadata: The gateway records the time of each request, the payload size, the API key, and the rate limit counters. Prem uses this metadata for billing, rate limits, and analytics. The metadata does not include your content. See the Security Model.
  • Error traces: Error responses include a support_id. Prem keeps the trace for that identifier for up to one week. See Errors.

​
Why you do not have to trust this page

Most providers publish a retention policy and ask you to trust it. Prem API adds two verifiable protections:
  1. The platform cannot read your content. Your device encrypts all content before it leaves. The gateway routes only encrypted bytes. Decryption happens only inside the sealed enclave. A stored copy of your traffic would be useless ciphertext. See Encryption.
  2. You can verify hardware evidence and freshness. An approved measurement policy is still required to prove that the expected release image is running. See Attestation.

​
Zero Data Retention mode

ZDR is the standard inference mode. Prem sends each ZDR request to a trusted partner. The partner operates under a zero-retention agreement. The partner keeps no prompt, completion, or audio content. The partner uses no content to train a model. The result for retention is the same. The assurance is different. In ZDR mode your content is plaintext at the Prem API Gateway and at the partner. There is no attestation evidence for you to examine. Read ZDR security boundary before you choose that mode for sensitive data.
Questions about retention, or requirements for a specific agreement? Contact us at support@premai.io.